Move the matrix stage of SkLinearBitmapPipeline over to using SkSmallAllocator.

The rest of the stages will follow. When all stages are completed,
this should significantly reduce stack use in the typical case.

This a step in removing the baroque stage system and moving towards
a SkRasterPipeline stage system.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2484273002

Review-Url: https://codereview.chromium.org/2484273002
3 files changed